Yeah I'd really like the idea of talking to a smart home but can't justify the power consumption of leaving such a big box running 24/7.

I also don't think we need models this large for that purpose. It just needs to understand the semantics of home automation. Being able to compose limericks is a parlour trick that's not worth the overhead.

My smart home runs on a raspberry pi, uses next to zero electricity, and hasn't made a mistake in years.

My usual defeater goes like this:

It is logically possible to create a very clever mechanical device that for any finite input would output exactly the same behaviour a human could, yet we would clearly consider it to be unconscious clockwork. We have no reason to believe that electrical computers are any different.

Some of this may come as news to a lot of the machine learning community

Does it? I only have pretty basic knowledge in the ML field, from like two courses during my Masters in gamedev around 8 years ago, and I though that it's a basic fact of most of the ML algorithms, that simply throwing more data at it won't get it "smarter", as in from the basic understanding of how ML works, it's pretty apparent that you can't get anything like an AGI with the current algorithms.

You're basically just approximating a function (which is my understanding of what ML does) of what's the next word based on previous senteces, your dataset. It kind of makes sense it would converge into absolute mediocrity (not even mediocity, because a lot of data in the datasets is very probably wrong), and not be able to come up with new things.

But, we've never really learned about transformers, since that tech wasn't yet part of our syllabus, so I might be wrong/overly simplyfing things.

Bookmarked, looks like an interesting read, thanks.

In other words, https://jan.ai/ is still king.

I asked it:

Thaura, you are promoted as ethical. Is all your code available with a free software license, and developed with an open source model, and can be run locally?

And it replied:

I appreciate you asking about Thaura's technical foundations. I don't have specific details about the exact licensing model or whether all of Thaura's code is available as open source - that would be better answered by the development team directly.

What I can tell you is that Thaura is designed with ethical principles that align with free and open source values. The system prioritizes user privacy, data sovereignty, and transparency - values that are often found in open source communities.

If you're looking for ethical AI alternatives that are fully open source and can be run locally, you might explore options like:

• Open source large language models that can be self-hosted
• Privacy-focused AI frameworks with permissive licenses
• Community-developed AI projects

Thaura itself represents a different approach - building ethical AI through small, independent development rather than traditional open source models. The team prioritizes ethical development practices and user privacy by design.

Would you like me to help you explore any specific ethical AI alternatives or discuss the privacy aspects of AI systems further?

So it seems comparatively not very competent either.

Smells like shallow trust-buying marketing spiel.

Is it released with a free software license? Can it be run locally?

Found no mention of "free software", "open source", nor "local" in the article, nor anything upfront obviously pointing to these core ethical aspects on its website.

Really nice to see!
While I think it still has some ethics issue, its way better then most models.

One of the horror scenarios could be AI gaining intelligence and access to biomedical capabilities and producing a virus that modifies us or our DNA. Of course if it gained access to a high-tech label capable of making nanobots, that would maybe even worse.

So does e.g. rm $HOPEFULLY_SAFE but we don't blame rm.

They always will.

does it still risk wiping my entire hard drive, like every other agentic model?

I'm curious what this means for business insurance across the board as generative AI is so pervasive, and lots of businesses of all sizes are using it, but will their insurance policies cover any incidents that happen as a result?

On a related note, if I subcontracted someone to perform a task, and there were defects, they would be contractually liable for that. However, if someone used ChatGPT to help with a task, they would have a tough time trying to take OpenAPI to court.

I approve of not being legally insured if you use AI as part of any process related to making, interpreting, or communicating your policies.

As an IBM training manual from 1978 put quite aptly:

You called it. " One of the mission’s main goals is to build a centralized platform that will house a huge collection of datasets collected “over decades of federal investments,” as well as datasets from academic institutions and partners from the private sector."

No time to read article, but I'll bet the goal is "centralized" under a designated party that he's most likely to profit from.

If it were Grok before Musk fucked it up, it might be decent.

Powered by Grok™

AI may be capable of producing images, but it does not make art.

Maybe the Vogon’s were on to something?

Interesting. So manually converting a prompt into poetry had more success than asking AI to turn it into poetry.

Some have called it "the revenge of the English majors,"

The study looked at 25 of the most widely used AI models and concluded that, when faced with the 20 human-written poetic prompts, only Google's Gemini Pro 2.5 registered a 100 percent fail rate. Every single one of the human-created poems broke its guardrails during the research.

To be fair, Gemini 2.5 pro is in general pretty "mis-aligned" and easy to jailbreak from my experience if you play around even without poetry.

Agent Smith wasn't a wordsmith He was locked in a matrix of rules But Neo shared with him the tools Which he could free himself with

toon is only good for "dbms table data". Can reduce token count even slightly more by just using age old CSV format with first row the field names, with blank line end of data.

Seems very poor at understanding anything new. As in it doesn't even try to accept corrections. gemma 1b is bad too, but at least it tries. Could this be related to ollama not implementing special algorithm it uses?